Mathematics Colloquium
Elizabeth Newman
Emory
Title: How to Train Your Network: Challenges and Opportunities
Date: Friday, September 27, 2024
Place and Time: Love 101, 3:05-3:55 pm
Abstract. Deep neural networks (DNNs) have become ubiquitous, revolutionizing computer vision and natural language processing and accelerating scientific simulations and discoveries. However, achieving these impressive results requires significant time and computational resources to learn a good DNN. The process of learning a DNN is called training and is typically posed as a challenging high-dimensional, non-convex, stochastic optimization problem. In this talk, we will review the basics of neural networks with significant focus on network training. We will then highlight how some old-school optimization strategies couples with modern hardware and software advancements offer promising approaches to training networks faster and more reliably.